Co-Optimization Two-Phase

Repository to reproduce the results of the paper "An Empirical Study on the Computation Budget of Co-Optimization of Robot Design and Control in Simulation"

Supported Frameworks

The project integrates with the following robotic simulation frameworks:

1. EvoGym - Soft robot co-design using PPO
2. RoboGrammar - Rigid robot design using MCTS and MPC
3. Tholiao - Modular soft robot simulation (not used in the paper)
4. GymREM2D - 2D modular robot environment
5. JorgenREM - Modular robot evolution
6. KevinCoAdapting - Co-adapting morphology and control (not used in the paper)

Requirements

• Python 3.7 (required for compatibility with some frameworks)
• CMake 3.20+
• Ubuntu 18.04 or later
• Various OpenGL libraries (see setup.sh)

Key Python Dependencies

• PyTorch 1.12.0
• NumPy, Pandas, Matplotlib
• Stable-Baselines3
• PyBullet
• scipy 1.4.1
• GPy, GPyOpt
• neat-python
• CMA-ES, Nevergrad
• See requirements.txt for complete list

Installation

Local Installation

1. Clone the repository:

git clone <repository-url>
cd co-optimization-two-phase

2. Run the setup script:

bash setup.sh --local

This will:

• Install system dependencies
• Create a Python 3.7 virtual environment
• Install Python packages
• Compile GLEW
• Build RoboGrammar with Python bindings
• Install EvoGym
• Set up GymREM2D with Box2D
• Install Revolve dependencies

3. Activate the environment:

source venv/bin/activate

Usage

Running Experiments

Each framework has its own experiment script in the src/ directory:

RoboGrammar Example

# Activate virtual environment
source venv/bin/activate

# Run experiment with index 0
python src/robogrammar_experiment.py --local_launch 0

# Generate plots
python src/robogrammar_experiment.py --plot

# Clean temporary files
python src/robogrammar_experiment.py --clean

EvoGym Example

python src/evogym_experiment.py --local_launch 0
python src/evogym_experiment.py --plot

Other Frameworks

python src/tholiao_experiment.py --local_launch 0
python src/gymrem2d_experiment.py --local_launch 0
python src/jorgenrem_experiment.py --local_launch 0
python src/kevincoadapting_experiment.py --local_launch 0

Cluster Execution

Submit jobs to the cluster using provided SLURM scripts:

sbatch cluster_scripts/launch_one_robogrammar.sl
sbatch cluster_scripts/launch_one_evogym.sl
sbatch cluster_scripts/launch_one_jorgenrem.sl

Generating Visualizations

Run all plots in background:

bash all_plots_background.sh

Configuration

The Parameters class in NestedOptimization.py (src/NestedOptimization.py:65) allows configuration of:

• Number of seeds for experiments
• Stopping criteria (max steps)
• Environment-specific parameters
• Inner optimization quantities and episode lengths

Results

Results are saved in results/<framework>/:

• data/ - Raw experimental data (.txt files)
• figures/ - Generated plots (.pdf files)
• videos/ - Robot animations (.mp4 files)

License

This software is released into the public domain under The Unlicense.

Note: The code in other_repos/ and glew-2.1.0/ directories are from external authors and may have different licenses. Please refer to each repository for their respective licenses.